ArrayMorph is a software to manage array data stored on cloud object storage efficiently. It supports both HDF5 C++ API and h5py API. The data returned by h5py API is numpy arrays. By using h5py API, users can access array data stored on the cloud and feed the read data into machine learning pipelines seamlessly.
The Camera Traps application is both a simulator and IoT device software for utilizing machine learning on the edge in field research. The first implementation specializes in applying computer vision (detection and classification) to wildlife images for animal ecology studies. Two operational modes are supported (1) an input dataset of images to act as the images that would be generated an IoT camera device or (2) an input video file that would be captured by a camera which is then processed by an image detecting plugin that saves frames with motion in them; it uses these images to drive the simulation.
The ctcontroller tool can be used to manage the provisioning and releasing of edge hardware as well as running and shutting down the camera-traps application.
GitHub Repo
A browser-based geospatial data discovery and collection interface for selecting an area of interest, checking satellite scene availability, configuring collection parameters, and submitting GeoHarmonizer collection jobs to HPC resources through Tapis.
How ArrayMorph Works
Architectual Overview
Architecture Overview
CKN facilitates seamless connectivity between edge devices and the cloud through event streaming, enabling real‑time data capture and processing. By leveraging event‑stream processing, it captures, aggregates, and stores historical system‑performance data in a knowledge graph that models application behaviour and guides model selection and deployment at the edge.
Area of Interest
Architecture
Architecture
Why use Ilúvatar for research?
What IEMS is
Convention and Usage
Software COMPonents
Available operations
Templates, versions, and runs
The Patra Toolkit embeds transparency and governance directly into the training workflow. Integrated scanners collect essential metadata—data sources, fairness metrics, and explainability insights—during model training and then generate a machine‑actionable JSON model card. These cards plug into the Patra Knowledge Base for rich queries on provenance, version history, and auditing. Flexible back‑ends publish models and artifacts to repositories such as Hugging Face or GitHub, automatically recording lineage links to trace every model’s evolution.
Patra Frontend is the web interface for the Patra AI Cards framework, giving users a way to browse, search, submit, and edit model cards and datasheets without calling the Patra Knowledge Base API directly.
Status Notice
Detailing csv columns
The core vision is to integrate a suite of expert-level compiler tools, enabling seamless, intelligent code analysis, optimization, and profiling for diverse programming languages.
Pipeline Architecture
FastKG is a knowledge graph embedding training library. Knowledge Graph (KG) embeddings are a way to represent entities and relationships from a KG in a continuous vector space, enabling tasks like link prediction and reasoning. TransE, a popular model, represents relationships as translations in the embedding space, such that for a valid triplet (head, relation, tail), the embedding of the head plus the relation vector is close to the embedding of the tail. Training data for TransE is typically stored in a tab-separated values (TSV) format, where each line represents a triplet, e.g., entity1\trelation1\tentity2. For example a dummy train.tsv should look like this:
Overview
Quick Start
Installation
WAYS to configure HARP to setup applications for profiling:
System Requirements
How-To Run on the ICICLE Instance
Install dependencies
See the full documentation for detailed instructions on creating custom plug‑ins and streaming events to the knowledge graph.
Run Earth Data Hub Locally
Installation
Authentication
Authentication
Try it Out
How to deploy a new control-plane release on Tapis Pods
The pipeline file (operations.json)
Sign in to Tapis
Installation
Hosted deployments (Tapis Pods)
Prerequisites
Setting up cetus
Install the project
How to Import Existing Annotations
FastAPI service that turns text into embedding vectors using Qwen3-Embedding-0.6B (GGUF quantized) via llama-cpp-python, designed for the ICICLE AI Tapis tenant. The service runs the model locally — no external API calls — so a single .gguf file plus a Tapis token is everything a deployment needs.
FastAPI + Qdrant vector storage and retrieval service for the ICICLE AI Tapis tenant. Clients provide their own pre-computed embeddings — the service handles storage, search, and reranking.
Ilúvatar is an open Serverless platform built with the goal of jumpstarting and streamlining FaaS research.
The Intelligent Edge Management Service (IEMS) is ICICLE's fleet control plane for deploying and operating AI workloads on edge devices. It provides a browser dashboard and REST API to enroll devices, publish model cards, deploy containers to Jetsons and similar hardware, monitor heartbeats, and manage live camera streams. The service integrates with Tapis for authentication and hosting, Patra for model provenance, and the ICICLE TapisUI extension for portal access.
iSpLib is an accelerated sparse kernel library with PyTorch interface. This library has an auto-tuner which generates optimized custom sparse kernels based on the user environment. The goal of this library is to provide efficient sparse operations for Graph Neural Network implementations. Currently it has support for CPU-based efficient Sparse Dense Matrix Multiplication (spmm-sum only) with autograd.
The ML Field Planner is a framework for analyzing ML pipelines and studying edge-to-center tradeoffs regarding function placement of ML. Using ML Field Planner, researchers configure experiments to run on real IoT hardware, configure machine learning models to analyze custom benchmark datasets, and experiment with different algorithm configurations, such as storage compression, all from a graphical user interface.
A browser-based OpenCV pipeline builder: build an image pre-processing
A browser-based workflow builder for the full ML lifecycle: wire up a pipeline in
GitHub Repo
Vue 3 + Vite web interface for the Patra Knowledge Base — the frontend component of the Patra AI Cards framework. Browse model cards and datasheets, submit and edit records, and explore edge-deployment experiments.
# Patra Knowledge Base
This project introduces an agentic approach for high-level and multi-purpose compilers
An AI-driven educational framework that integrates automated curriculum generation with an end-to-end computer vision pipeline, enabling learning for high school and undergraduate students combining domain agnostic datasets with machine learning and AI concepts.
A 7-step HPC-backed pipeline for few-shot object detection — from interactive image annotation through class support generation, proposal visualization, and Tapis-powered job execution — built to work across any research domain. Also includes a dedicated Semantic Segmentation pipeline for polygon and pixel-mask annotation with SAM3-assisted labeling.
Cetus Source to Source compiler improvements are being done at the University of Delaware. In this release, we release cetus base plus an static profiling feature and the instructions to use it.
SpMM Example
Set up Environment Variables
Run a simple example: Writing and Reading HDF5 files from Cloud
Create a Custom CKN Plug-in
Open Earth Data Hub
Quickstart
Quickstart
Tutorial: Run the control plane locally
Build a pipeline (editor)
Build a workflow (canvas)
Building a Patra Model Card
The Smart Compiler leverages AI models alongside traditional compiler techniques to boost the performance and scalability of C and Python programs. Through intelligent profiling and optimization strategies, it identifies and applies enhancements to improve code efficiency.
This documentation goes over how to use Smart Curriculum Designer, the machine learning and AI curriculum generator. The application itself enables educators to generate models, content, exercises, and solutions, weaving the domain/dataset specified by the educator.
Running Your First Detection Pipeline